fix(speech): forward a top-level ttl and pass deployment headers as is through the TTS bridge

Picked from 9dafd85341 without its awaited aspeech path. 1.108.x has no internal sub-call scope around the bridge's inner chat call, so awaiting acompletion there writes the speech spend row under the provider's response id instead of the call id

litellm/endpoints/speech/speech_to_completion_bridge/handler.py

@@ -90,7 +90,7 @@ class SpeechToCompletionBridgeHandler:
         logging_obj: "LiteLLMLoggingObj",
         custom_llm_provider: str,
         api_base: str | None,
-        extra_headers: Mapping[str, str] | None,
+        extra_headers: Mapping[str, object] | None,
         timeout: "float | httpx.Timeout | None",
         cache_controls: "CacheControlsHint",
     ) -> "HttpxBinaryResponseContent":

litellm/endpoints/speech/speech_to_completion_bridge/transformation.py

@@ -88,7 +88,7 @@ class SpeechToCompletionBridgeTransformationHandler:
         headers: dict,
         litellm_logging_obj: "LiteLLMLoggingObj",
         custom_llm_provider: str,
-    ) -> dict:
+    ) -> Mapping[str, object]:
         self._validate_response_format(model, custom_llm_provider, optional_params)
         user_message: Final[ChatCompletionUserMessage] = {"role": "user", "content": input}
         return_kwargs: Final = {

litellm/types/utils.py

@@ -4874,7 +4874,8 @@ class BaseUrlRoutingHint(TypedDict, total=False):
 class CacheControlsHint(TypedDict, total=False):
     cache: ReadOnly[OnErrorOmit[Mapping[str, object] | None]]
     caching: ReadOnly[OnErrorOmit[JsonValue]]
+    ttl: ReadOnly[OnErrorOmit[JsonValue]]
 
 
 class DeploymentHeadersHint(TypedDict, total=False):
-    extra_headers: ReadOnly[OnErrorOmit[Mapping[str, str] | None]]
+    extra_headers: ReadOnly[OnErrorOmit[Mapping[str, object] | None]]

tests/unit/test_main.py

@@ -36,6 +36,7 @@ from litellm import (
 from litellm import main as litellm_main
 from litellm.caching.base_cache import BaseCache
 from litellm.caching.caching import Cache
+from litellm.caching.in_memory_cache import InMemoryCache
 from litellm.constants import CONTROL_OPTIONS_KEY
 from litellm.integrations.custom_logger import CustomLogger
 from litellm.integrations.custom_prompt_management import CustomPromptManagement
@@ -4905,7 +4906,7 @@ def test_speech_gemini_tts_honors_the_request_cache_controls(
 
 @pytest.mark.parametrize(
     "malformed_controls",
-    [{"caching": "yes"}, {"caching": 1}, {"cache": {1: True}}, {"extra_headers": {"x-gateway-key": 5}}],
+    [{"caching": "yes"}, {"caching": 1}, {"cache": {1: True}}],
 )
 def test_speech_gemini_tts_without_a_cache_ignores_malformed_controls(
     respx_mock: respx.MockRouter,
@@ -4929,6 +4930,61 @@ def test_speech_gemini_tts_without_a_cache_ignores_malformed_controls(
     assert response.content.startswith(b"RIFF")
 
 
+@pytest.mark.parametrize("ttl", [60, "60"])
+def test_speech_gemini_tts_keeps_cached_audio_only_for_the_request_ttl(
+    respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch, ttl: int | str
+):
+    now: Final = [0.0]
+    cache: Final = Cache(type="local")
+    cache.cache = InMemoryCache(clock=lambda: now[0])
+    monkeypatch.setattr(litellm, "cache", cache)
+    route: Final = respx_mock.post(url__regex=GEMINI_DEFAULT_GENERATE_CONTENT_URL).mock(
+        return_value=httpx.Response(200, json=_gemini_tts_generate_content_response())
+    )
+
+    def speak() -> None:
+        litellm.speech(
+            model=f"gemini/{GEMINI_TTS_MODEL}", input="hello", voice="Kore", api_key="fake-gemini-key", ttl=ttl
+        )
+
+    speak()
+    now[0] = 30.0
+    speak()
+    assert route.call_count == 1
+    now[0] = 61.0
+    speak()
+    assert route.call_count == 2
+
+
+def test_speech_gemini_tts_fails_like_chat_on_a_non_string_deployment_header(respx_mock: respx.MockRouter):
+    route: Final = respx_mock.post(url__regex=GEMINI_DEFAULT_GENERATE_CONTENT_URL).mock(
+        return_value=httpx.Response(200, json=_gemini_tts_generate_content_response())
+    )
+    deployment_headers: Final = {"x-gateway-key": "gateway-secret", "x-org-id": 12345}
+
+    with pytest.raises(litellm.APIConnectionError) as chat_error:
+        litellm.completion(
+            model=f"gemini/{GEMINI_TTS_MODEL}",
+            messages=[{"role": "user", "content": "hello"}],
+            modalities=["audio"],
+            audio={"voice": "Kore", "format": "pcm16"},
+            api_key="fake-gemini-key",
+            extra_headers=deployment_headers,
+        )
+    with pytest.raises(litellm.APIConnectionError) as speech_error:
+        litellm.speech(
+            model=f"gemini/{GEMINI_TTS_MODEL}",
+            input="hello",
+            voice="Kore",
+            api_key="fake-gemini-key",
+            extra_headers=deployment_headers,
+        )
+
+    assert "Header value must be str or bytes" in str(chat_error.value)
+    assert "Header value must be str or bytes" in str(speech_error.value)
+    assert route.call_count == 0
+
+
 FOUNDRY_HOST: Final = "https://my-project.services.ai.azure.com"